VASSNet: A Vessel-Aware Appearance-Structure Synergistic Network with Hierarchical Semantic Optimization for Retinal Vessel Segmentation

Abstract

Retinal vessel segmentation underpins risk assessment for ocular and systemic diseases. Existing encoder-decoder networks remain constrained by domain-specific appearance shifts, weak responses from small vessels, and cross-level semantic misalignment, leading to missed distal branches, topological discontinuities, and background false positives. To address these limitations, we propose a vessel-aware appearance-structure synergistic network (VASSNet) with layer-aware semantics. Channel-specific style augmentation (CSA) models channel-dependent responses and perturbs appearance statistics to reduce reliance on domain-specific color distributions. Structure-adaptive topological masking (SAT-Mask) exploits vessel skeletons, local scales, and topological critical points to simulate small-vessel attenuation, promoting continuous reconstruction from incomplete cues. Semantic-guided adaptive context (SGAC) reformulates skip fusion through semantic calibration, multi-scale context aggregation, and residual enhancement, strengthening the representation of vessel trunks, branches, and boundaries. Experiments on DRIVE, STARE, and CHASE_DB1 demonstrate that VASSNet achieves the highest AUC across all three datasets while maintaining strong small-vessel detection and suppressing complex background interference. Cross-domain, ablation, and qualitative analyses further substantiate its generalization capability and the effectiveness of each component.